Shanghai AI Lab unveils a science discovery platform built to close the loop from model to wet lab
量子位 · wechat · 2026-07-21
Shanghai AI Lab’s science platform ties together models, agents, and wet labs
This long article presents “Shusheng·Duanyan”, a scientific discovery platform from the Shanghai AI Lab, and argues that it aims to turn AI into a full research loop rather than a mere assistant.
Core idea
- The platform tries to solve the “impossible triangle” in brain science: complex cognition, human-like models, and long, expensive experiments.
- In zebrafish research, it builds a virtual fish with anatomy-grounded constraints and a closed-loop brain–body–environment simulation.
- The article claims the digital zebrafish can predict behavior with 69%+ accuracy, while AI scientist agents match expert conclusions at 75%+.
Platform architecture
It describes five linked layers:
- Compute via DeepLink,
- Data via a scientific database with 28M+ papers,
- Model via a 397B science model,
- Agents for task decomposition,
- Autonomous experiments for dry-to-wet validation.
Reported results
The platform is said to have helped across several areas, including:
- brain science,
- materials and electroplating,
- quantum computing,
- aerospace design,
- and earth system modeling.
The broader pitch
The author frames the platform as a move from “repeating the known” to producing the unknown, with safety, provenance, and reproducibility built into the workflow.
More from Research
- Stanford Data-Centric AI Workshop reopens site, extends poster deadline — jure · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Ito: a browser that detects browsing intent and auto-organizes tabs into Flows — andrewkuznet · 2026-10-26
- Cornell study says brains process language as linear sequences, challenging Chomsky's hierarchy — anirbanbandyo · 2026-10-26
- Hi3D V1.0 open-sourced with 2M training assets and 90% slicing success rate — kimmonismus · 2026-10-26
- Ortet evaluates 34 open-weight models on 50 health tasks, finds general leaderboards insufficient — elmanmansimov · 2026-10-26